1
PRESENTED BY
Kanav Mathur
2427030777
School of Computer Science and Engineering
Department of Computer Science and Engineering
Federated Spatio-Temporal Graph
Network
SUPERVISED BY
Dr. Shishir Singh Chauhan
2
03
04
Introduction
05
06
Literature Review
07 Problem Statement
08 Objectives
09 Proposed Solution
10
11
12
Outline
3
Introduction
03
Urban traffic congestion is a significant global challenge
Real-time traffic forecasting enhances:
•
Route planning
•
Signal Control
•
Emergency Response
Traffic data obtained from a network of road sensors
Conventional deep learning models rely on centralized
data
Real-world scenarios are geographically distributed
4
Federated Learning for
Traffic
Traffic sensors are geographically divided
Data ownership and privacy issues
Communication limitations between regions
Centralized models can:
•
Fail to generalize well
•
Disregard regional heterogeneity
Require:
A distributed learning framework that:
•
Respects data privacy
•
Captures spatial-temporal patterns
•
Generalizes well to new city regions
This Photo by Unknown Author is licensed under CC BY-SA
5
Spatio-Temporal
Traffic Models
Graph-based Models:
STGCN (Spatio-Temporal Graph Convolutional Networks)
DCRNN
STGAT (Graph Attention Networks)
Key Ideas:
Use road graph to model spatial dependency
Use convolution or RNN to model temporal dynamics
Use adjacency matrix to represent traffic influence propagation
Limitations:
Need centralized data access
Assume fixed graph structure
6
Federated Learning in
Intelligent Systems
Federated Learning(FL):
Local training on distributed clients
Aggregation using FedAvg
No raw data sharing
Applications:
Healthcare
Mobile systems
IoT networks
Gap in Literature:
Few studies on federated graph learning for traffic
Few studies on inductive generalization to unseen regions
7
Problem Statement
How can
1. Different zones in city learn traffic patterns without sharing data but still
understand roads causing traffic
2. Global model generalize to unseen regions
8
Objectives
Primary Objectives
Design centralized STGCN baseline
Implement federated traffic learning simulation
Test inductive generalization
Implement adaptive graph learning
Performance Objectives
Achieve minimum MAE and RMSE
Compare:
Centralized vs federated
Fixed graph vs adaptive graph
Test robustness across regions
9
Proposed Solution
Key Components:
•
Local STGCN per region
•
Adaptive graph correction module
•
Federated parameter aggregation
•
Inductive evaluation on unseen region
Core Novelty:
•
Learn spatial dependencies
collaboratively
•
Preserve regional data privacy
10
Dataset and Preprocessing
Dataset: METR-LA
207 traffic sensors
5-minute intervals
Road distance-based adjacency
Speed values are stored
Sensors are pre-ided
Preprocessing:
Missing value imputation
Z-score normalization
Sliding window generation (12 12) →
11
Sensor Clustering & Federated Simulation
12
Local STGCN Architecture
13
Learnable Spatial Dependency Modelling
Benefits:
Captures hidden traffic correlations
Models indirect congestion propagation
Improves generalization across regions
14
Federated Optimization Procedure
Advantages:
Privacy preservation
Reduced communication overhead
Scalable architecture
15
Purpose:
Test real-world deployability
Evaluate cross-region knowledge transfer
Demonstrate inductive capability
Generalization to Unseen Region
16
Training Details & Optimization
Loss Function Mean Squared Error(MSE)
Evaluation Metrics MAE,RMSE
Optimizer Adam
Learning Rate 0.001
Batch Size 32
Epochs 10-20 per round
Activation ReLU
Regularization Early stopping
Weight sharing across client
17
Result
Outcome
I aim to achieve drafting the research paper for this model.